At Dust, the Paris-based platform that facilitates collaboration between humans and AI agents in the workplace, internships for 15- and 16-year-old high school students have served as an unintended testing ground since 2022. On June 17, at the opening of the tenth edition of VivaTech, its co-founder and CTO, Stanislas Polu, discussed what four successive internship programs reveal about the transfer of knowledge, at a time when execution is delegated but judgment is not.
The Apprentice and the Artifact
The intern is fifteen years old. He has just presented to the entire team the video game he designed all by himself in a week: a three-dimensional world, a character that walks, and physics that work. Everyone congratulates him and asks him about the character model—how did he create it? He explains, with the ease of a seasoned pro, that an agent “downloaded something from somewhere on the internet.” They press him: can he show them the code? The boy looks at them, genuinely perplexed. “The what?” He had been running a chatbot for five days without ever opening the file it was generating.
Stanislas Polu recounts this anecdote without any apparent concern. Since 2022, his company has welcomed one to three high school students, aged 15 or 16, each year for one- or two-week internships, with computers set up to use the same tools as the employees. Progress, he says, has come in distinct stages. Three years ago, an intern with no programming experience taught himself the basics of JavaScript by querying a natural language model—which was already remarkable for the time. Two years ago, another intern translated an entire set of technical documentation from English to French, then delivered a functional implementation on his own, still fully aware of what he was writing. A year ago, the intern no longer looked at the code: he was directing, in everyday language, an agent that wrote it for him—and only discovered, when asked by a colleague, that this code, as an object, existed.
This shift can be summed up in an observation by Polu: What mattered a year ago was being reliable and fast; what matters now is knowing what to build—a capacity for judgment that is hard to acquire without first having failed, started over, and understood the material with one’s own hands.
What the workshop already knew
This question is nothing new to anyone who has observed a traditional crafts workshop. The value of a master craftsman has never been measured by how quickly he performs a task, but by his ability to sense—before the piece is finished—that it will not be quite right. This skill cannot be taught quickly: it requires years of trial and error, of starting over, and of learning the material’s resistance. Chatbots promise the opposite—a result without the friction that, in craftsmanship, is precisely what shapes one’s judgment.
Polu himself is hesitant to make a call. When asked what his former intern will one day need to understand about the code he had him write without reading it, he simply replies that he has no idea. This admission is just as telling as the anecdote: the person who designs these agents for thousands of companies doesn’t know whether the skills they’re designed to eliminate will one day become necessary again.
Details. On November 24, 2025, Anthropic released Claude Opus 4.5 with a two-thirds reduction in usage costs compared to its predecessor—a threshold that the Dust teams describe as the catalyst for a lasting shift: the proportion of code still written by hand by developers within the company dropped in just a few months from a large majority to about one-fifth.
Five days before the opening of VivaTech on June 12, 2026, the U.S. government ordered Anthropic to suspend access to its most capable models— Fable 5 and Mythos 5—a retroactive and unprecedented decision, made just four days after their commercial launch. On stage, Polu described the era as dizzying rather than alarming. It remains to be seen whether the speed with which a fifteen-year-old high school student learns to control something he doesn’t understand says anything truly new about expertise—or whether it merely reveals, sooner than expected, what every apprentice has always had to learn on his own: deciding, before taking action, what is worth doing.
Cette publication est également disponible en :
